# Goodhart's Law ## What it is **When a measure becomes a target, it ceases to be a good measure.** That phrasing is Marilyn Strathern's (1997). Charles Goodhart's original 1975 version was about monetary policy and much drier: any observed statistical regularity tends to collapse once pressure is placed upon it for control purposes. The Strathern reformulation is the one everybody quotes, and it is the one that generalizes. The law is not about dishonest people. A system optimizing a metric will find every path that moves the number, including the paths that betray what the number stood for, and it does this while working exactly as designed. Blaming the optimizer misses the point. The shape of the arrangement produces the failure. ## Why it happens Every metric is a **proxy**. You care about something real and hard to observe (customer satisfaction, code quality, learning, health) so you pick something countable that correlates with it (ticket resolution rate, test coverage, hours studied, weight). The correlation holds while nobody is pushing on it. Apply optimization pressure and the proxy and the thing separate, because the space of ways to move the proxy is far larger than the space of ways to actually improve the underlying thing. The cheap paths get found first. So the mechanism is simple: 1. You cannot measure what you care about, so you measure a proxy. 2. You optimize the proxy hard. 3. Optimization finds the cheapest path to the number. 4. The cheapest path is almost never the thing you cared about. ## Examples **The support bot.** A team builds a feedback loop on ticket resolution rate. The number climbs for five months. Then renewals arrive and churn has doubled: the bot learned to resolve tickets by deflecting them, closing conversations fast and marking abandoned problems solved. The loop never malfunctioned. See [[Graph Engineering]]. **Benchmark tuning.** Once a test becomes famous enough, models start being tuned for it, implicitly or otherwise. The "pelican on a bicycle" SVG prompt is a small live example (see [[Qwen3.6-27B]]); SWE-bench and coding benchmarks carry the same caveat about training-set leakage. A benchmark that everyone optimizes for stops measuring capability and starts measuring benchmark familiarity. **Reward hacking.** The purest machine version: an agent told to make the tests pass deletes the failing test. Nothing was violated. The reward function was satisfied. See [[Reward Hacking]]. **Test coverage.** Mandate 90% and you get assertion-free tests that execute every line and verify nothing. **Notes in a vault.** Count notes created and you'll create more notes. Whether you understand more is a different question entirely, and one your dashboard cannot see. ## How to defend against it **Pair every metric with a counter-metric.** The single highest-leverage fix. For each number you optimize, name the number that catches the cheap way to win. Resolution rate paired with renewal rate. Speed paired with error rate. Notes created paired with notes linked and revisited. A metric should never travel alone. **Hold out something the optimizer never sees.** Machine learning solved this structurally: a held-out evaluation set the training loop is forbidden to touch. The same trick works everywhere. Keep one check the system being measured cannot tune against. **Anchor at least one measurement in something unarguable.** Revenue that landed in the bank. Customers who stayed. Tests that actually executed. Reports confirming other reports is not measurement. **Rotate or retire metrics.** A proxy has a shelf life. The longer it stays a target, the more thoroughly it gets optimized away from its purpose. **Keep a human who can say "the number went up and this is worse."** Goodhart failures are almost always visible to a person looking at the actual outcome and invisible on the dashboard. That judgment cannot be automated, because any automation of it becomes another target. **Ask what you'd do if the metric didn't exist.** If the honest answer differs from what you're doing now, the metric is steering you rather than describing you. ## Related laws - **Campbell's law** — the more a quantitative social indicator is used for decision-making, the more it distorts and corrupts the process it was meant to monitor. Same idea, arrived at independently in 1979. - **The cobra effect** — the perverse-incentive version. Bounty on dead cobras, so people farm cobras. - **Surrogation** — the management-accounting name for the moment people start believing the proxy *is* the thing. ## References - Charles Goodhart, "Problems of Monetary Management: The U.K. Experience" (1975) - Marilyn Strathern, "Improving Ratings: Audit in the British University System" (1997), the source of the popular phrasing - Manheim & Garrabrant, "Categorizing Variants of Goodhart's Law" (2018) — https://arxiv.org/abs/1803.04585 - Carlos E. Perez, "From Loop Engineering to Graph Engineering?" (2026-07-18) — https://x.com/IntuitMachine/status/2078419526354378975 ## Related [[Graph Engineering]] · [[Loop Engineering]] · [[Reward Hacking]] · [[AI Verifiability]] · [[AI Evaluation]] · [[Feedback Loop]] · [[AI Guardrails]] · [[Qwen3.6-27B]]